Adversarial Defense 벤치마크
Adversarial Defense on ImageNet (non-targeted PGD, max perturbation=4)
Accuracy
- 2019-04-29 — ResNet-152 free-m=4: Accuracy 36.0
- 2019-07-04 — LLR-ResNet-152: Accuracy 47.0
- 2020-06-25 — SAT-EfficientNet-L1: Accuracy 58.6
| Rank | Model | Accuracy | Paper | Code | Year |
|---|---|---|---|---|---|
| 1 | SAT-EfficientNet-L1 | 58.6% | Smooth Adversarial Training | cihangxie/SmoothAdversarialTraining | 2020 |
| 2 | LLR-ResNet-152 | 47.0% | Adversarial Robustness through Local Linearization | 2019 | |
| 3 | ResNet-152 free-m=4 | 36.0% | Adversarial Training for Free! | locuslab/fast_adversarial · mahyarnajibi/FreeAdversarialTraining · ashafahi/free_adv_train · +3 | 2019 |
| 4 | ResNet-101 free-m=4 | 34.3% | Adversarial Training for Free! | locuslab/fast_adversarial · mahyarnajibi/FreeAdversarialTraining · ashafahi/free_adv_train · +3 | 2019 |
| 5 | ResNet-50 free-m=4 | 31.8% | Adversarial Training for Free! | locuslab/fast_adversarial · mahyarnajibi/FreeAdversarialTraining · ashafahi/free_adv_train · +3 | 2019 |